Causal Deep Learning with Applications in Healthcare

Abstract

I am interested in inferring causality from observational data. Such causal inference is organised in two main categories: (i) causal discovery, and (ii) causal effect inference. With (i) we aim to determine whether or not a drug is causally responsible for an outcome; while (ii) already assumes the drug is responsible, but infers the extent to which the drug is responsible. My proposals to these inference problems stem from a new field: causal deep learning, which I introduce in chapter 2. Both categories are important in science, where this dissertation considers applications in healthcare in particular. In a clinical setting, both of these problems are typically addressed using a randomised clinical trial (RCT), where treatments are assigned randomly to patients. While very powerful, RCTs are not always feasible, hence our aim is to answer the same questions without randomising treatment. This dissertation is organised as follows: chapter 2, introduces Causal Deep Learning, the general framework which we’ll use throughout the text. In chapter 3 we will solve two problems in causal discovery: a first on structural causal discovery, and a second on structural equation discovery in a temporal setting. Based on structural causal information, we continue in chapter 5 with causal treatment effect inference where we are particularly interested in handling missing values in treatment effects datasets. Equipped with causal inference methods, we will apply treatment effects models in the practical setting of organ transplantation, with a specific focus on donor liver allocation in chapter 6. Finally, we will discuss evaluating such allocation systems, using underlying causal structures in chapter 7. As such, this dissertation advances causality on both a methodological level (in both discovery as well as inference), as well as a practical level with a specific focus on healthcare.